diff --git a/3rdparty/spirv-cross/spirv_common.hpp b/3rdparty/spirv-cross/spirv_common.hpp
index 0cdbda23d..51fffe7a1 100644
--- a/3rdparty/spirv-cross/spirv_common.hpp
+++ b/3rdparty/spirv-cross/spirv_common.hpp
@@ -220,7 +220,7 @@ static inline std::string convert_to_string(int32_t value)
 	// INT_MIN is ... special on some backends. If we use a decimal literal, and negate it, we
 	// could accidentally promote the literal to long first, then negate.
 	// To workaround it, emit int(0x80000000) instead.
-	if (value == std::numeric_limits<int32_t>::min())
+	if (value == (std::numeric_limits<int32_t>::min)())
 		return "int(0x80000000)";
 	else
 		return std::to_string(value);
@@ -231,7 +231,7 @@ static inline std::string convert_to_string(int64_t value, const std::string &in
 	// INT64_MIN is ... special on some backends.
 	// If we use a decimal literal, and negate it, we might overflow the representable numbers.
 	// To workaround it, emit int(0x80000000) instead.
-	if (value == std::numeric_limits<int64_t>::min())
+	if (value == (std::numeric_limits<int64_t>::min)())
 		return join(int64_type, "(0x8000000000000000u", (long_long_literal_suffix ? "ll" : "l"), ")");
 	else
 		return std::to_string(value) + (long_long_literal_suffix ? "ll" : "l");
diff --git a/3rdparty/spirv-cross/spirv_cross.cpp b/3rdparty/spirv-cross/spirv_cross.cpp
index 4ab985efd..3f0fec334 100644
--- a/3rdparty/spirv-cross/spirv_cross.cpp
+++ b/3rdparty/spirv-cross/spirv_cross.cpp
@@ -640,6 +640,13 @@ bool Compiler::is_physical_pointer(const SPIRType &type) const
 	return type.op == OpTypePointer && type.storage == StorageClassPhysicalStorageBuffer;
 }
 
+bool Compiler::is_physical_pointer_to_buffer_block(const SPIRType &type) const
+{
+	return is_physical_pointer(type) && get_pointee_type(type).self == type.parent_type &&
+	       (has_decoration(type.self, DecorationBlock) ||
+	        has_decoration(type.self, DecorationBufferBlock));
+}
+
 bool Compiler::is_runtime_size_array(const SPIRType &type)
 {
 	return type.op == OpTypeRuntimeArray;
@@ -5024,8 +5031,7 @@ void Compiler::PhysicalStorageBufferPointerHandler::mark_aligned_access(uint32_t
 bool Compiler::PhysicalStorageBufferPointerHandler::type_is_bda_block_entry(uint32_t type_id) const
 {
 	auto &type = compiler.get<SPIRType>(type_id);
-	return type.storage == StorageClassPhysicalStorageBufferEXT && type.pointer &&
-	       type.pointer_depth == 1 && !compiler.type_is_array_of_pointers(type);
+	return compiler.is_physical_pointer(type);
 }
 
 uint32_t Compiler::PhysicalStorageBufferPointerHandler::get_minimum_scalar_alignment(const SPIRType &type) const
@@ -5055,7 +5061,8 @@ void Compiler::PhysicalStorageBufferPointerHandler::setup_meta_chain(uint32_t ty
 		access_chain_to_physical_block[var_id] = &meta;
 
 		auto &type = compiler.get<SPIRType>(type_id);
-		if (type.basetype != SPIRType::Struct)
+
+		if (!compiler.is_physical_pointer_to_buffer_block(type))
 			non_block_types.insert(type_id);
 
 		if (meta.alignment == 0)
@@ -5114,9 +5121,7 @@ bool Compiler::PhysicalStorageBufferPointerHandler::handle(Op op, const uint32_t
 uint32_t Compiler::PhysicalStorageBufferPointerHandler::get_base_non_block_type_id(uint32_t type_id) const
 {
 	auto *type = &compiler.get<SPIRType>(type_id);
-	while (type->pointer &&
-	       type->storage == StorageClassPhysicalStorageBufferEXT &&
-	       !type_is_bda_block_entry(type_id))
+	while (compiler.is_physical_pointer(*type) && !type_is_bda_block_entry(type_id))
 	{
 		type_id = type->parent_type;
 		type = &compiler.get<SPIRType>(type_id);
@@ -5131,12 +5136,10 @@ void Compiler::PhysicalStorageBufferPointerHandler::analyze_non_block_types_from
 	for (auto &member : type.member_types)
 	{
 		auto &subtype = compiler.get<SPIRType>(member);
-		if (subtype.basetype != SPIRType::Struct && subtype.pointer &&
-		    subtype.storage == spv::StorageClassPhysicalStorageBufferEXT)
-		{
+
+		if (compiler.is_physical_pointer(subtype) && !compiler.is_physical_pointer_to_buffer_block(subtype))
 			non_block_types.insert(get_base_non_block_type_id(member));
-		}
-		else if (subtype.basetype == SPIRType::Struct && !subtype.pointer)
+		else if (subtype.basetype == SPIRType::Struct && !compiler.is_pointer(subtype))
 			analyze_non_block_types_from_block(subtype);
 	}
 }
@@ -5149,9 +5152,14 @@ void Compiler::analyze_non_block_pointer_types()
 	// Analyze any block declaration we have to make. It might contain
 	// physical pointers to POD types which we never used, and thus never added to the list.
 	// We'll need to add those pointer types to the set of types we declare.
-	ir.for_each_typed_id<SPIRType>([&](uint32_t, SPIRType &type) {
-		if (has_decoration(type.self, DecorationBlock) || has_decoration(type.self, DecorationBufferBlock))
+	ir.for_each_typed_id<SPIRType>([&](uint32_t id, SPIRType &type) {
+		// Only analyze the raw block struct, not any pointer-to-struct, since that's just redundant.
+		if (type.self == id &&
+		    (has_decoration(type.self, DecorationBlock) ||
+		     has_decoration(type.self, DecorationBufferBlock)))
+		{
 			handler.analyze_non_block_types_from_block(type);
+		}
 	});
 
 	physical_storage_non_block_pointer_types.reserve(handler.non_block_types.size());
diff --git a/3rdparty/spirv-cross/spirv_cross.hpp b/3rdparty/spirv-cross/spirv_cross.hpp
index 9fe6c41c4..2ba602502 100644
--- a/3rdparty/spirv-cross/spirv_cross.hpp
+++ b/3rdparty/spirv-cross/spirv_cross.hpp
@@ -685,6 +685,7 @@ protected:
 	bool is_array(const SPIRType &type) const;
 	bool is_pointer(const SPIRType &type) const;
 	bool is_physical_pointer(const SPIRType &type) const;
+	bool is_physical_pointer_to_buffer_block(const SPIRType &type) const;
 	static bool is_runtime_size_array(const SPIRType &type);
 	uint32_t expression_type_id(uint32_t id) const;
 	const SPIRType &expression_type(uint32_t id) const;
diff --git a/3rdparty/spirv-cross/spirv_cross_containers.hpp b/3rdparty/spirv-cross/spirv_cross_containers.hpp
index e79b32093..c496cb75b 100644
--- a/3rdparty/spirv-cross/spirv_cross_containers.hpp
+++ b/3rdparty/spirv-cross/spirv_cross_containers.hpp
@@ -576,6 +576,7 @@ public:
 			if (!ptr)
 				return nullptr;
 
+			vacants.reserve(num_objects);
 			for (unsigned i = 0; i < num_objects; i++)
 				vacants.push_back(&ptr[i]);
 
diff --git a/3rdparty/spirv-cross/spirv_cross_error_handling.hpp b/3rdparty/spirv-cross/spirv_cross_error_handling.hpp
index e96ebb9a7..91e6cf4f8 100644
--- a/3rdparty/spirv-cross/spirv_cross_error_handling.hpp
+++ b/3rdparty/spirv-cross/spirv_cross_error_handling.hpp
@@ -66,6 +66,11 @@ public:
 	    : std::runtime_error(str)
 	{
 	}
+
+	explicit CompilerError(const char *str)
+	    : std::runtime_error(str)
+	{
+	}
 };
 
 #define SPIRV_CROSS_THROW(x) throw CompilerError(x)
diff --git a/3rdparty/spirv-cross/spirv_glsl.cpp b/3rdparty/spirv-cross/spirv_glsl.cpp
index 0efc44aee..0d37ba2e8 100644
--- a/3rdparty/spirv-cross/spirv_glsl.cpp
+++ b/3rdparty/spirv-cross/spirv_glsl.cpp
@@ -1768,7 +1768,9 @@ bool CompilerGLSL::buffer_is_packing_standard(const SPIRType &type, BufferPackin
 	for (uint32_t i = 0; i < type.member_types.size(); i++)
 	{
 		auto &memb_type = get<SPIRType>(type.member_types[i]);
-		auto member_flags = ir.meta[type.self].members[i].decoration_flags;
+
+		auto *type_meta = ir.find_meta(type.self);
+		auto member_flags = type_meta ? type_meta->members[i].decoration_flags : Bitset{};
 
 		// Verify alignment rules.
 		uint32_t packed_alignment = type_to_packed_alignment(memb_type, member_flags, packing);
@@ -2145,11 +2147,11 @@ string CompilerGLSL::layout_for_variable(const SPIRVariable &var)
 	// If SPIR-V does not comply with either layout, we cannot really work around it.
 	if (can_use_buffer_blocks && (ubo_block || emulated_ubo))
 	{
-		attr.push_back(buffer_to_packing_standard(type, false));
+		attr.push_back(buffer_to_packing_standard(type, false, true));
 	}
 	else if (can_use_buffer_blocks && (push_constant_block || ssbo_block))
 	{
-		attr.push_back(buffer_to_packing_standard(type, true));
+		attr.push_back(buffer_to_packing_standard(type, true, true));
 	}
 
 	// For images, the type itself adds a layout qualifer.
@@ -2170,7 +2172,9 @@ string CompilerGLSL::layout_for_variable(const SPIRVariable &var)
 	return res;
 }
 
-string CompilerGLSL::buffer_to_packing_standard(const SPIRType &type, bool support_std430_without_scalar_layout)
+string CompilerGLSL::buffer_to_packing_standard(const SPIRType &type,
+                                                bool support_std430_without_scalar_layout,
+                                                bool support_enhanced_layouts)
 {
 	if (support_std430_without_scalar_layout && buffer_is_packing_standard(type, BufferPackingStd430))
 		return "std430";
@@ -2182,6 +2186,7 @@ string CompilerGLSL::buffer_to_packing_standard(const SPIRType &type, bool suppo
 		return "scalar";
 	}
 	else if (support_std430_without_scalar_layout &&
+	         support_enhanced_layouts &&
 	         buffer_is_packing_standard(type, BufferPackingStd430EnhancedLayout))
 	{
 		if (options.es && !options.vulkan_semantics)
@@ -2193,7 +2198,8 @@ string CompilerGLSL::buffer_to_packing_standard(const SPIRType &type, bool suppo
 		set_extended_decoration(type.self, SPIRVCrossDecorationExplicitOffset);
 		return "std430";
 	}
-	else if (buffer_is_packing_standard(type, BufferPackingStd140EnhancedLayout))
+	else if (support_enhanced_layouts &&
+	         buffer_is_packing_standard(type, BufferPackingStd140EnhancedLayout))
 	{
 		// Fallback time. We might be able to use the ARB_enhanced_layouts to deal with this difference,
 		// however, we can only use layout(offset) on the block itself, not any substructs, so the substructs better be the appropriate layout.
@@ -2207,7 +2213,9 @@ string CompilerGLSL::buffer_to_packing_standard(const SPIRType &type, bool suppo
 		set_extended_decoration(type.self, SPIRVCrossDecorationExplicitOffset);
 		return "std140";
 	}
-	else if (options.vulkan_semantics && buffer_is_packing_standard(type, BufferPackingScalarEnhancedLayout))
+	else if (options.vulkan_semantics &&
+	         support_enhanced_layouts &&
+	         buffer_is_packing_standard(type, BufferPackingScalarEnhancedLayout))
 	{
 		set_extended_decoration(type.self, SPIRVCrossDecorationExplicitOffset);
 		require_extension_internal("GL_EXT_scalar_block_layout");
@@ -2221,6 +2229,7 @@ string CompilerGLSL::buffer_to_packing_standard(const SPIRType &type, bool suppo
 		return "std430";
 	}
 	else if (!support_std430_without_scalar_layout && options.vulkan_semantics &&
+	         support_enhanced_layouts &&
 	         buffer_is_packing_standard(type, BufferPackingStd430EnhancedLayout))
 	{
 		// UBOs can support std430 with GL_EXT_scalar_block_layout.
@@ -2319,7 +2328,7 @@ void CompilerGLSL::emit_buffer_reference_block(uint32_t type_id, bool forward_de
 	auto &type = get<SPIRType>(type_id);
 	string buffer_name;
 
-	if (forward_declaration)
+	if (forward_declaration && is_physical_pointer_to_buffer_block(type))
 	{
 		// Block names should never alias, but from HLSL input they kind of can because block types are reused for UAVs ...
 		// Allow aliased name since we might be declaring the block twice. Once with buffer reference (forward declared) and one proper declaration.
@@ -2352,10 +2361,10 @@ void CompilerGLSL::emit_buffer_reference_block(uint32_t type_id, bool forward_de
 		// Ensure we emit the correct name when emitting non-forward pointer type.
 		ir.meta[type.self].decoration.alias = buffer_name;
 	}
-	else if (type.basetype != SPIRType::Struct)
-		buffer_name = type_to_glsl(type);
 	else
-		buffer_name = to_name(type.self, false);
+	{
+		buffer_name = type_to_glsl(type);
+	}
 
 	if (!forward_declaration)
 	{
@@ -2364,13 +2373,13 @@ void CompilerGLSL::emit_buffer_reference_block(uint32_t type_id, bool forward_de
 		if (itr != physical_storage_type_to_alignment.end())
 			alignment = itr->second.alignment;
 
-		if (type.basetype == SPIRType::Struct)
+		if (is_physical_pointer_to_buffer_block(type))
 		{
 			SmallVector<std::string> attributes;
 			attributes.push_back("buffer_reference");
 			if (alignment)
 				attributes.push_back(join("buffer_reference_align = ", alignment));
-			attributes.push_back(buffer_to_packing_standard(type, true));
+			attributes.push_back(buffer_to_packing_standard(type, true, true));
 
 			auto flags = ir.get_buffer_block_type_flags(type);
 			string decorations;
@@ -2385,14 +2394,32 @@ void CompilerGLSL::emit_buffer_reference_block(uint32_t type_id, bool forward_de
 
 			statement("layout(", merge(attributes), ")", decorations, " buffer ", buffer_name);
 		}
-		else if (alignment)
-			statement("layout(buffer_reference, buffer_reference_align = ", alignment, ") buffer ", buffer_name);
 		else
-			statement("layout(buffer_reference) buffer ", buffer_name);
+		{
+			string packing_standard;
+			if (type.basetype == SPIRType::Struct)
+			{
+				// The non-block type is embedded in a block, so we cannot use enhanced layouts :(
+				packing_standard = buffer_to_packing_standard(type, true, false) + ", ";
+			}
+			else if (is_array(get_pointee_type(type)))
+			{
+				SPIRType wrap_type{OpTypeStruct};
+				wrap_type.self = ir.increase_bound_by(1);
+				wrap_type.member_types.push_back(get_pointee_type_id(type_id));
+				ir.set_member_decoration(wrap_type.self, 0, DecorationOffset, 0);
+				packing_standard = buffer_to_packing_standard(wrap_type, true, false) + ", ";
+			}
+
+			if (alignment)
+				statement("layout(", packing_standard, "buffer_reference, buffer_reference_align = ", alignment, ") buffer ", buffer_name);
+			else
+				statement("layout(", packing_standard, "buffer_reference) buffer ", buffer_name);
+		}
 
 		begin_scope();
 
-		if (type.basetype == SPIRType::Struct)
+		if (is_physical_pointer_to_buffer_block(type))
 		{
 			type.member_name_cache.clear();
 
@@ -3705,31 +3732,34 @@ void CompilerGLSL::emit_resources()
 
 	if (ir.addressing_model == AddressingModelPhysicalStorageBuffer64EXT)
 	{
-		for (auto type : physical_storage_non_block_pointer_types)
-		{
-			emit_buffer_reference_block(type, false);
-		}
-
 		// Output buffer reference blocks.
 		// Do this in two stages, one with forward declaration,
 		// and one without. Buffer reference blocks can reference themselves
 		// to support things like linked lists.
-		ir.for_each_typed_id<SPIRType>([&](uint32_t self, SPIRType &type) {
-			if (type.basetype == SPIRType::Struct && type.pointer &&
-			    type.pointer_depth == 1 && !type_is_array_of_pointers(type) &&
-			    type.storage == StorageClassPhysicalStorageBufferEXT)
+		ir.for_each_typed_id<SPIRType>([&](uint32_t id, SPIRType &type) {
+			if (is_physical_pointer(type))
 			{
-				emit_buffer_reference_block(self, true);
+				bool emit_type = true;
+				if (!is_physical_pointer_to_buffer_block(type))
+				{
+					// Only forward-declare if we intend to emit it in the non_block_pointer types.
+					// Otherwise, these are just "benign" pointer types that exist as a result of access chains.
+					emit_type = std::find(physical_storage_non_block_pointer_types.begin(),
+					                      physical_storage_non_block_pointer_types.end(),
+					                      id) != physical_storage_non_block_pointer_types.end();
+				}
+
+				if (emit_type)
+					emit_buffer_reference_block(id, true);
 			}
 		});
 
-		ir.for_each_typed_id<SPIRType>([&](uint32_t self, SPIRType &type) {
-			if (type.basetype == SPIRType::Struct &&
-			    type.pointer && type.pointer_depth == 1 && !type_is_array_of_pointers(type) &&
-			    type.storage == StorageClassPhysicalStorageBufferEXT)
-			{
-				emit_buffer_reference_block(self, false);
-			}
+		for (auto type : physical_storage_non_block_pointer_types)
+			emit_buffer_reference_block(type, false);
+
+		ir.for_each_typed_id<SPIRType>([&](uint32_t id, SPIRType &type) {
+			if (is_physical_pointer_to_buffer_block(type))
+				emit_buffer_reference_block(id, false);
 		});
 	}
 
@@ -5011,11 +5041,8 @@ string CompilerGLSL::dereference_expression(const SPIRType &expr_type, const std
 		return expr.substr(1);
 	else if (backend.native_pointers)
 		return join('*', expr);
-	else if (expr_type.storage == StorageClassPhysicalStorageBufferEXT && expr_type.basetype != SPIRType::Struct &&
-	         expr_type.pointer_depth == 1)
-	{
+	else if (is_physical_pointer(expr_type) && !is_physical_pointer_to_buffer_block(expr_type))
 		return join(enclose_expression(expr), ".value");
-	}
 	else
 		return expr;
 }
@@ -15695,17 +15722,29 @@ string CompilerGLSL::type_to_glsl_constructor(const SPIRType &type)
 // depend on a specific object's use of that type.
 string CompilerGLSL::type_to_glsl(const SPIRType &type, uint32_t id)
 {
-	if (type.pointer && type.storage == StorageClassPhysicalStorageBufferEXT && type.basetype != SPIRType::Struct)
+	if (is_physical_pointer(type) && !is_physical_pointer_to_buffer_block(type))
 	{
 		// Need to create a magic type name which compacts the entire type information.
-		string name = type_to_glsl(get_pointee_type(type));
-		for (size_t i = 0; i < type.array.size(); i++)
+		auto *parent = &get_pointee_type(type);
+		string name = type_to_glsl(*parent);
+
+		uint32_t array_stride = get_decoration(type.parent_type, DecorationArrayStride);
+
+		// Resolve all array dimensions in one go since once we lose the pointer type,
+		// array information is left to to_array_type_glsl. The base type loses array information.
+		while (is_array(*parent))
 		{
-			if (type.array_size_literal[i])
-				name += join(type.array[i], "_");
+			if (parent->array_size_literal.back())
+				name += join(type.array.back(), "_");
 			else
-				name += join("id", type.array[i], "_");
+				name += join("id", type.array.back(), "_");
+
+			name += "stride_" + std::to_string(array_stride);
+
+			array_stride = get_decoration(parent->parent_type, DecorationArrayStride);
+			parent = &get<SPIRType>(parent->parent_type);
 		}
+
 		name += "Pointer";
 		return name;
 	}
diff --git a/3rdparty/spirv-cross/spirv_glsl.hpp b/3rdparty/spirv-cross/spirv_glsl.hpp
index 4cba0bb63..3ce044e39 100644
--- a/3rdparty/spirv-cross/spirv_glsl.hpp
+++ b/3rdparty/spirv-cross/spirv_glsl.hpp
@@ -833,7 +833,9 @@ protected:
 	bool buffer_is_packing_standard(const SPIRType &type, BufferPackingStandard packing,
 	                                uint32_t *failed_index = nullptr, uint32_t start_offset = 0,
 	                                uint32_t end_offset = ~(0u));
-	std::string buffer_to_packing_standard(const SPIRType &type, bool support_std430_without_scalar_layout);
+	std::string buffer_to_packing_standard(const SPIRType &type,
+	                                       bool support_std430_without_scalar_layout,
+	                                       bool support_enhanced_layouts);
 
 	uint32_t type_to_packed_base_size(const SPIRType &type, BufferPackingStandard packing);
 	uint32_t type_to_packed_alignment(const SPIRType &type, const Bitset &flags, BufferPackingStandard packing);
diff --git a/3rdparty/spirv-cross/spirv_msl.cpp b/3rdparty/spirv-cross/spirv_msl.cpp
index de81738e7..e73751d7f 100644
--- a/3rdparty/spirv-cross/spirv_msl.cpp
+++ b/3rdparty/spirv-cross/spirv_msl.cpp
@@ -1938,10 +1938,14 @@ void CompilerMSL::extract_global_variables_from_function(uint32_t func_id, std::
 				// When using the pointer, we need to know which variable it is actually loaded from.
 				uint32_t base_id = ops[2];
 				auto *var = maybe_get_backing_variable(base_id);
-				if (var && atomic_image_vars_emulated.count(var->self))
+				if (var)
 				{
-					if (!get<SPIRType>(var->basetype).array.empty())
-						SPIRV_CROSS_THROW("Cannot emulate array of storage images with atomics. Use MSL 3.1 for native support.");
+					if (atomic_image_vars_emulated.count(var->self) &&
+					    !get<SPIRType>(var->basetype).array.empty())
+					{
+						SPIRV_CROSS_THROW(
+								"Cannot emulate array of storage images with atomics. Use MSL 3.1 for native support.");
+					}
 
 					if (global_var_ids.find(base_id) != global_var_ids.end())
 						added_arg_ids.insert(base_id);
@@ -7412,19 +7416,28 @@ void CompilerMSL::emit_custom_functions()
 			break;
 
 		case SPVFuncImplVariableDescriptorArray:
-			statement("template<typename T>");
-			statement("struct spvDescriptorArray");
-			begin_scope();
-			statement("spvDescriptorArray(const device spvDescriptor<T>* ptr) : ptr(ptr)");
-			begin_scope();
-			end_scope();
-			statement("const device T& operator [] (size_t i) const");
-			begin_scope();
-			statement("return ptr[i].value;");
-			end_scope();
-			statement("const device spvDescriptor<T>* ptr;");
-			end_scope_decl();
-			statement("");
+			if (spv_function_implementations.count(SPVFuncImplVariableDescriptor) != 0)
+			{
+				statement("template<typename T>");
+				statement("struct spvDescriptorArray");
+				begin_scope();
+				statement("spvDescriptorArray(const device spvDescriptor<T>* ptr) : ptr(ptr)");
+				begin_scope();
+				end_scope();
+				statement("const device T& operator [] (size_t i) const");
+				begin_scope();
+				statement("return ptr[i].value;");
+				end_scope();
+				statement("const device spvDescriptor<T>* ptr;");
+				end_scope_decl();
+				statement("");
+			}
+			else
+			{
+				statement("template<typename T>");
+				statement("struct spvDescriptorArray;");
+				statement("");
+			}
 
 			if (msl_options.runtime_array_rich_descriptor &&
 			    spv_function_implementations.count(SPVFuncImplVariableSizedDescriptor) != 0)
@@ -7458,6 +7471,22 @@ void CompilerMSL::emit_custom_functions()
 			statement("");
 			break;
 
+		case SPVFuncImplReduceAdd:
+			// Metal doesn't support __builtin_reduce_add or simd_reduce_add, so we need this.
+			// Metal also doesn't support the other vector builtins, which would have been useful to make this a single template.
+
+			statement("template <typename T>");
+			statement("T reduce_add(vec<T, 2> v) { return v.x + v.y; }");
+
+			statement("template <typename T>");
+			statement("T reduce_add(vec<T, 3> v) { return v.x + v.y + v.z; }");
+
+			statement("template <typename T>");
+			statement("T reduce_add(vec<T, 4> v) { return v.x + v.y + v.z + v.w; }");
+
+			statement("");
+			break;
+
 		default:
 			break;
 		}
@@ -9641,6 +9670,132 @@ void CompilerMSL::emit_instruction(const Instruction &instruction)
 		break;
 	}
 
+	case OpSDot:
+	case OpUDot:
+	case OpSUDot:
+	{
+		uint32_t result_type = ops[0];
+		uint32_t id = ops[1];
+		uint32_t vec1 = ops[2];
+		uint32_t vec2 = ops[3];
+
+		auto &input_type1 = expression_type(vec1);
+		auto &input_type2 = expression_type(vec2);
+
+		string vec1input, vec2input;
+		auto input_size = input_type1.vecsize;
+		if (instruction.length == 5)
+		{
+			if (ops[4] == PackedVectorFormatPackedVectorFormat4x8Bit)
+			{
+				string type = opcode == OpSDot || opcode == OpSUDot ? "char4" : "uchar4";
+				vec1input = join("as_type<", type, ">(", to_expression(vec1), ")");
+				type = opcode == OpSDot ? "char4" : "uchar4";
+				vec2input = join("as_type<", type, ">(", to_expression(vec2), ")");
+				input_size = 4;
+			}
+			else
+				SPIRV_CROSS_THROW("Packed vector formats other than 4x8Bit for integer dot product is not supported.");
+		}
+		else
+		{
+			// Inputs are sign or zero-extended to their target width.
+			SPIRType::BaseType vec1_expected_type =
+					opcode != OpUDot ?
+					to_signed_basetype(input_type1.width) :
+					to_unsigned_basetype(input_type1.width);
+
+			SPIRType::BaseType vec2_expected_type =
+					opcode != OpSDot ?
+					to_unsigned_basetype(input_type2.width) :
+					to_signed_basetype(input_type2.width);
+
+			vec1input = bitcast_expression(vec1_expected_type, vec1);
+			vec2input = bitcast_expression(vec2_expected_type, vec2);
+		}
+
+		auto &type = get<SPIRType>(result_type);
+
+		// We'll get the appropriate sign-extend or zero-extend, no matter which type we cast to here.
+		// The addition in reduce_add is sign-invariant.
+		auto result_type_cast = join(type_to_glsl(type), input_size);
+
+		string exp = join("reduce_add(",
+		                  result_type_cast, "(", vec1input, ") * ",
+		                  result_type_cast, "(", vec2input, "))");
+
+		emit_op(result_type, id, exp, should_forward(vec1) && should_forward(vec2));
+		inherit_expression_dependencies(id, vec1);
+		inherit_expression_dependencies(id, vec2);
+		break;
+	}
+
+	case OpSDotAccSat:
+	case OpUDotAccSat:
+	case OpSUDotAccSat:
+	{
+		uint32_t result_type = ops[0];
+		uint32_t id = ops[1];
+		uint32_t vec1 = ops[2];
+		uint32_t vec2 = ops[3];
+		uint32_t acc = ops[4];
+
+		auto input_type1 = expression_type(vec1);
+		auto input_type2 = expression_type(vec2);
+
+		string vec1input, vec2input;
+		if (instruction.length == 6)
+		{
+			if (ops[5] == PackedVectorFormatPackedVectorFormat4x8Bit)
+			{
+				string type = opcode == OpSDotAccSat || opcode == OpSUDotAccSat ? "char4" : "uchar4";
+				vec1input = join("as_type<", type, ">(", to_expression(vec1), ")");
+				type = opcode == OpSDotAccSat ? "char4" : "uchar4";
+				vec2input = join("as_type<", type, ">(", to_expression(vec2), ")");
+				input_type1.vecsize = 4;
+				input_type2.vecsize = 4;
+			}
+			else
+				SPIRV_CROSS_THROW("Packed vector formats other than 4x8Bit for integer dot product is not supported.");
+		}
+		else
+		{
+			// Inputs are sign or zero-extended to their target width.
+			SPIRType::BaseType vec1_expected_type =
+					opcode != OpUDotAccSat ?
+					to_signed_basetype(input_type1.width) :
+					to_unsigned_basetype(input_type1.width);
+
+			SPIRType::BaseType vec2_expected_type =
+					opcode != OpSDotAccSat ?
+					to_unsigned_basetype(input_type2.width) :
+					to_signed_basetype(input_type2.width);
+
+			vec1input = bitcast_expression(vec1_expected_type, vec1);
+			vec2input = bitcast_expression(vec2_expected_type, vec2);
+		}
+
+		auto &type = get<SPIRType>(result_type);
+
+		SPIRType::BaseType pre_saturate_type =
+				opcode != OpUDotAccSat ?
+				to_signed_basetype(type.width) :
+				to_unsigned_basetype(type.width);
+
+		input_type1.basetype = pre_saturate_type;
+		input_type2.basetype = pre_saturate_type;
+
+		string exp = join(type_to_glsl(type), "(addsat(reduce_add(",
+		                  type_to_glsl(input_type1), "(", vec1input, ") * ",
+		                  type_to_glsl(input_type2), "(", vec2input, ")), ",
+						  bitcast_expression(pre_saturate_type, acc), "))");
+
+		emit_op(result_type, id, exp, should_forward(vec1) && should_forward(vec2));
+		inherit_expression_dependencies(id, vec1);
+		inherit_expression_dependencies(id, vec2);
+		break;
+	}
+
 	default:
 		CompilerGLSL::emit_instruction(instruction);
 		break;
@@ -10052,6 +10207,9 @@ void CompilerMSL::emit_atomic_func_op(uint32_t result_type, uint32_t result_id,
 		// Emulate texture2D atomic operations
 		if (res_type.storage == StorageClassUniformConstant && res_type.basetype == SPIRType::Image)
 		{
+			auto &flags = ir.get_decoration_bitset(var->self);
+			if (decoration_flags_signal_volatile(flags))
+				exp += "volatile ";
 			exp += "device";
 		}
 		else
@@ -12708,6 +12866,11 @@ string CompilerMSL::get_argument_address_space(const SPIRVariable &argument)
 	return get_type_address_space(type, argument.self, true);
 }
 
+bool CompilerMSL::decoration_flags_signal_volatile(const Bitset &flags)
+{
+	return flags.get(DecorationVolatile) || flags.get(DecorationCoherent);
+}
+
 string CompilerMSL::get_type_address_space(const SPIRType &type, uint32_t id, bool argument)
 {
 	// This can be called for variable pointer contexts as well, so be very careful about which method we choose.
@@ -12817,7 +12980,7 @@ string CompilerMSL::get_type_address_space(const SPIRType &type, uint32_t id, bo
 		addr_space = type.pointer || (argument && type.basetype == SPIRType::ControlPointArray) ? "thread" : "";
 	}
 
-	return join(flags.get(DecorationVolatile) || flags.get(DecorationCoherent) ? "volatile " : "", addr_space);
+	return join(decoration_flags_signal_volatile(flags) ? "volatile " : "", addr_space);
 }
 
 const char *CompilerMSL::to_restrict(uint32_t id, bool space)
@@ -13551,7 +13714,9 @@ void CompilerMSL::entry_point_args_discrete_descriptors(string &ep_args)
 			// Emulate texture2D atomic operations
 			if (atomic_image_vars_emulated.count(var.self))
 			{
-				ep_args += ", device atomic_" + type_to_glsl(get<SPIRType>(basetype.image.type), 0);
+				auto &flags = ir.get_decoration_bitset(var.self);
+				const char *cv_flags = decoration_flags_signal_volatile(flags) ? "volatile " : "";
+				ep_args += join(", ", cv_flags, "device atomic_", type_to_glsl(get<SPIRType>(basetype.image.type), 0));
 				ep_args += "* " + r.name + "_atomic";
 				ep_args += " [[buffer(" + convert_to_string(r.secondary_index) + ")";
 				if (interlocked_resources.count(var_id))
@@ -14646,7 +14811,9 @@ string CompilerMSL::argument_decl(const SPIRFunction::Parameter &arg)
 	auto *backing_var = maybe_get_backing_variable(name_id);
 	if (backing_var && atomic_image_vars_emulated.count(backing_var->self))
 	{
-		decl += ", device atomic_" + type_to_glsl(get<SPIRType>(var_type.image.type), 0);
+		auto &flags = ir.get_decoration_bitset(backing_var->self);
+		const char *cv_flags = decoration_flags_signal_volatile(flags) ? "volatile " : "";
+		decl += join(", ", cv_flags, "device atomic_", type_to_glsl(get<SPIRType>(var_type.image.type), 0));
 		decl += "* " + to_expression(name_id) + "_atomic";
 	}
 
@@ -17266,6 +17433,14 @@ CompilerMSL::SPVFuncImpl CompilerMSL::OpCodePreprocessor::get_spv_func_impl(Op o
 	case OpGroupNonUniformQuadSwap:
 		return SPVFuncImplQuadSwap;
 
+	case OpSDot:
+	case OpUDot:
+	case OpSUDot:
+	case OpSDotAccSat:
+	case OpUDotAccSat:
+	case OpSUDotAccSat:
+		return SPVFuncImplReduceAdd;
+
 	default:
 		break;
 	}
diff --git a/3rdparty/spirv-cross/spirv_msl.hpp b/3rdparty/spirv-cross/spirv_msl.hpp
index 06744b873..ed9180972 100644
--- a/3rdparty/spirv-cross/spirv_msl.hpp
+++ b/3rdparty/spirv-cross/spirv_msl.hpp
@@ -823,7 +823,8 @@ protected:
 		SPVFuncImplVariableDescriptor,
 		SPVFuncImplVariableSizedDescriptor,
 		SPVFuncImplVariableDescriptorArray,
-		SPVFuncImplPaddedStd140
+		SPVFuncImplPaddedStd140,
+		SPVFuncImplReduceAdd
 	};
 
 	// If the underlying resource has been used for comparison then duplicate loads of that resource must be too
@@ -1044,6 +1045,7 @@ protected:
 	bool validate_member_packing_rules_msl(const SPIRType &type, uint32_t index) const;
 	std::string get_argument_address_space(const SPIRVariable &argument);
 	std::string get_type_address_space(const SPIRType &type, uint32_t id, bool argument = false);
+	static bool decoration_flags_signal_volatile(const Bitset &flags);
 	const char *to_restrict(uint32_t id, bool space);
 	SPIRType &get_stage_in_struct_type();
 	SPIRType &get_stage_out_struct_type();